fix(megdnn): fix megdnn benchmark testcase

GitOrigin-RevId: 726971474a8b056b7bb42004ae1a89a054562ae6

fix(megdnn): fix megdnn benchmark testcase
GitOrigin-RevId: 726971474a8b056b7bb42004ae1a89a054562ae6
97beae2f · Megvii Engine Team · 1d4baa42 · 97beae2f · 97beae2f
隐藏空白更改
内联并排

Showing with 12 addition and 4 deletion

dnn/test/cuda/batch_conv_bias.cpp dnn/test/cuda/batch_conv_bias.cpp +8 -4

dnn/test/cuda/benchmark.cpp dnn/test/cuda/benchmark.cpp +4 -0

未找到文件。
--- a/dnn/test/cuda/batch_conv_bias.cpp
+++ b/dnn/test/cuda/batch_conv_bias.cpp
@@ -241,10 +241,14 @@ void benchmark_target_algo(Handle* handle, const std::vector<BenchArgs>& args,
    "v" V(CUDNN_MAJOR) "." V(CUDNN_MINOR) "." V(CUDNN_PATCHLEVEL)
    benchmarker_cudnn.set_before_exec_callback(
            conv_bias::ConvBiasAlgoChecker<ConvBiasForward>(
-                    "CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_"
+                    ConvBiasForward::algo_name<ConvBias::DefaultParam>(
-                    "GEMM" CUDNN_VERSION_STRING));
+                            "CUDNN:ConvBiasActivation:"
-    benchmarker_matmul.set_before_exec_callback(
+                            "CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_"
-            AlgoChecker<BatchedMatrixMul>("BRUTE_FORCE-CUBLAS"));
+                            "GEMM" CUDNN_VERSION_STRING,
+                            {})
+                            .c_str()));
+    benchmarker_matmul.set_before_exec_callback(AlgoChecker<BatchedMatrixMul>(
+            ExecutionPolicyAlgoName{"BRUTE_FORCE", {{"CUBLAS", {}}}}));
    benchmarker.set_dtype(0, src_dtype)
            .set_dtype(1, filter_dtype)

--- a/dnn/test/cuda/benchmark.cpp
+++ b/dnn/test/cuda/benchmark.cpp
@@ -41,10 +41,12 @@ TEST_F(CUDA, BENCHMARK_CONVOLUTION_8X8X32)
        auto time_in_ms_float = benchmarker.set_param(param_float)
            .set_dtype(0, dtype::Float32())
            .set_dtype(1, dtype::Float32())
+            .set_dtype(2, dtype::Float32())
            .execs({src_float, filter_float, {}});
        auto time_in_ms_int = benchmarker.set_param(param_int)
            .set_dtype(0, dtype::Int8())
            .set_dtype(1, dtype::Int8())
+            .set_dtype(2, dtype::Int32())
            .execs({src_int, filter_int, {}});
        std::cout << "1x1: N=" << N << " OC=" << OC << " IC=" << IC
            << " H=" << H << " W=" << W
@@ -67,10 +69,12 @@ TEST_F(CUDA, BENCHMARK_CONVOLUTION_8X8X32)
        auto time_in_ms_float = benchmarker.set_param(param_float)
            .set_dtype(0, dtype::Float32())
            .set_dtype(1, dtype::Float32())
+            .set_dtype(2, dtype::Float32())
            .execs({src_float, filter_float, {}});
        auto time_in_ms_int = benchmarker.set_param(param_int)
            .set_dtype(0, dtype::Int8())
            .set_dtype(1, dtype::Int8())
+            .set_dtype(2, dtype::Int32())
            .execs({src_int, filter_int, {}});
        std::cout << "chanwise: N=" << N << " C=" << C
            << " H=" << H << " W=" << W << " F=" << F